ROBUST AND ADAPTIVE TECHNIQUES IN SELF-ORGANIZING NEURAL NETWORKS

Robust and adaptive training algorithms aiming at enhancing the capabi lities of self-organizing and Radial Basis Function (RBF) neural netwo rks are reviewed in this paper. The following robust variants of Learn ing Vector Quantizer (LVQ) are described: the order statistics LVQ, th e L-2 LVQ and the split-merge LVQ. Successful application of the margi nal median LVQ that belongs to the class of order statistics LVQs in t he self-organized selection of the centers in RBF neural networks is r eported. Moreover, the use of the median absolute deviation in the est imation of the covariance matrix of the observations assigned to each hidden unit in RBF neural networks is proposed. Applications that prov e the superiority of the proposed variants of LVQ and RBF neural netwo rks in noisy color image segmentation, color-based image recognition, segmentation of ultrasonic images, motion-field smoothing and moving o bject segmentation are outlined.